Low-Shot multi-label incremental learning for thoracic diseases diagnosis

Despite promising results of 14 types of diseases continuously reported on the large-scale NIH dataset, the applicability on real clinical practice with the deep learning based CADx for chest X-ray may still be quite elusive. It is because tens of diseases can be found in the chest X-ray and require to keep on learning and diagnosis. In this paper, we propose a low-shot multi-label incremental learning framework involving three phases, i.e., representation learning, low-shot learning and all-label fine-tuning phase, to demonstrate the feasibility and practicality of thoracic disease abnormalities of CADx in clinic. To facilitate the incremental learning in new small dataset situation, we also formulate a feature regularization prior, say multi-label squared gradient magnitude (MLSGM) to ensure the generalization capability of the deep learning model. The proposed approach has been evaluated on the public ChestX-ray14 dataset covering 14 types of basic abnormalities and a new small dataset MyX-ray including 6 types of novel abnormalities collected from Mianyang Central Hospital. The experimental result shows MLSGM method improves the average Area-Under-Curve (AUC) score on 6 types of novel abnormalities up to 7.6 points above the baseline when shot number is only 10. With the low-shot multi-label incremental learning framework, the AI application for the reading and diagnosis of chest X-ray over-all diseases and abnormalities can be possibly realized in clinic practice.